Impish_QWEN_14B-1M: A Naughty and Playful Qwen Finetune

Impish_QWEN_14B-1M, developed by SicariusSicariiStuff, is a 14.8 billion parameter model based on the Qwen2.5-14B-Instruct-1M architecture. This model stands out for its supreme context length of one million tokens, making it highly suitable for extended interactions and complex narratives. It is the first finetune of the Qwen2.5-14B-Instruct-1M to be released.

Key Capabilities & Features

• Extended Context: Boasts a 1 million token context window, enabling deep and prolonged conversations or creative writing tasks.
• Roleplay Optimization: Specifically trained for strong roleplay, catering to internet RP formats with medium-sized paragraphs.
• Low Censorship: Designed with a very low censorship level, offering a "naughty and playful" persona without requiring jailbreaks.
• High IFEval Score: Achieves a notable IFEval (0-Shot) score of 78.68, indicating strong instruction following for a 14B roleplay model.
• Improved Benchmarks: Reportedly has higher benchmarks than the original Qwen2.5-14B-Instruct-1M finetune.
• Experimental Data: Incorporates experimental data to neutralize GPT's inherent ideological biases.

Good For

• Role-Play: Excels in character-driven interactions and narrative generation.
• Creative Writing: Ideal for generating imaginative content and stories.
• General Tasks: Capable of handling a variety of general language tasks, leveraging its Qwen intelligence.

Users are advised to utilize the provided recommended settings for both assistant and roleplay modes to optimize performance, including specific repetition_penalty ranges and the SICAtxt format for character and adventure descriptions.